Recent advances in large language models (LLMs) have shown remarkable performance in various tasks, but whether and how LLMs could help with fake news detection remains underexplored.

The gap

Consensus across the literature

Supporting evidence — 3 representative gaps

• CONSISTENCY-BASED EVALUATION OF SHAP AND LIME EXPLANATIONS FOR MACHINE LEARNING-BASED FAKE NEWS DETECTION (2026) · doi

  This study provided a consistency-based evaluation of SHAP and LIME explanations for fake news detection using machine learning. The primary aim was not just to create a robust fake news classifier, but also to investigate whether two widely used explanation techniques offer consistent and reliable explanations for the same model predictions. This is crucial as fake news detection is a sensitive task and the user requires more than just a predicted label. They also require evidence that is fully trustworthy and understandable in support of the stated prediction. The WELFake dataset, which includes fake and real news samples at the article level [15] [16] was used for the experiment. The final cleaned data set comprised of 63,547 articles. Stratified splitting was used to split the dataset into training, validation and test sets. The machine learning pipeline used was TF-IDF based and multiple models were tested. The final model chosen was a balanced Logistic Regression classifier with 80,000 unigram TF-IDF features and a decision threshold of 0.52 that was optimized on the validation set. The chosen model had a good classification accuracy on the test set. It obtained 96.06% pg. 159 KJMR VOL.03 NO. 05 (2026) CONSISTENCY-BASED EVALUATION OF SHAP AND LIME ……. accuracy, 95.65% F1-score, 99.31% ROC-AUC, 99.18% PR-AUC, and 92.05% MCC. The confusion matrix also revealed a balanced distribution of errors, with 253 fake articles being classified as real and 248 real articles being classified as fake. From these results, it can be concluded that the trained classifier is reliable enough for further explanation analysis. The main contribution of the study is the corrected SHAP-LIME explanation comparison. The implementation aligns both explanation methods to the same predicted class before comparison. This is important because, in binary classification, SHAP values may naturally describe the direction of class 1, while LIME explains the predicted class. Without this correction, the comparison can become unfair and misleading. The study also normalizes explanation features by lowercasing, removing punctuation, normalizing spaces, and merging duplicate feature forms. This makes SHAP and LIME explanations more directly comparable at the word level. The corrected SHAP-LIME explanation comparison is the main contribution of the study. The implementation matches both explanation approaches to the same anticipated class prior to the comparison. This is crucial since in binary classification, SHAP values can naturally describe the direction of class 1, while LIME explains the predicted class. If this correction is not made, the comparison can be completely unfair and misguiding. The study also performs normalization of explanation features, which includes lowercasing,

  Keywords: explanation shap lime fake comparison class news used predicted based explanations classifier model real articles
• Bad Actor, Good Advisor: Exploring the Role of Large Language Models in Fake News Detection (2024) · doi

  Recent advances in large language models (LLMs) have shown remarkable performance in various tasks, but whether and how LLMs could help with fake news detection remains underexplored.

  Keywords: llms recent advances large language models remarkable performance various tasks whether help fake news detection
• Political Fake News Detection Farmwork in Social Networks Utilizing Hybrid Deep Learning Algorithms (2026) · doi

  In recent years, many researchers have investigated various machine learning algorithms to detect fake news, but they are often limited by their own characteristics such as dealing with big data, redundant features and lack of consideration of semantic and context information.

  Keywords: recent years researchers investigated various machine learning algorithms detect fake news often limited characteristics dealing